Intermittent Pneumatic Compression

Intermittent pneumatic compression (IPC) systems are generally used as a prophylactic for venous insufficiency (e.g., deep vein thrombosis) in the lower limbs by emulating the action of the muscle pump. Recent studies have demonstrated that compressing and releasing the calf within each heartbeat enhances blood flow compared to typical “slow” compression methods. However, each individual’s hemodynamics can be sensitive to the compression timing within the heartbeat. Assessing the optimal compression timing is a complex process, since it changes from one individual to another, but also for every heartbeat. Therefore, a good understanding of the relationship between the hemodynamics and compression timing is required. The aim of this study is to assess whether blood flow in the femoral artery is predictable when external pressure is applied to the calf at different times within the cardiac cycle. Four participants wore a custom IPC device and the timing of the compression within the cardiac cycle was randomly varied for one hour. The following measurements were collected: femoral blood velocity (BV), electrocardiogram (ECG), and the applied pressure. Predicting the BV is a two-step process. (1) ECG is predicted one sample ahead by a nonlinear auto-regressive (NAR) model. (2) The femoral blood velocity is predicted by a NAR with exogenous inputs (NARX) model using the ECG and the measured external pressure as external inputs. Our NAR and NARX models consist of artificial neural network models that were trained to predict the blood flow one sample ahead (~10 ms). Once trained, those models were used in a closed-loop form to predict the ECG and the blood velocity (BV) traces of the next heartbeat. Even though the models failed to predict the next ECG R-wave, the prediction of the ECG and the BV was accurate between two successive R-waves, showing that the BV can be predicted for one heartbeat ahead. Keywords-component; Intermittent pneumatic compression, deep vein thrombosis, blood flow, machine learning